Machine Vision and Applications
A fully automated framework for lung tumour detection, segmentation and analysis
--Manuscript Draft--
Manuscript Number:
Full Title: A fully automated framework for lung tumour detection, segmentation and analysis
Article Type: Full Paper
Corresponding Author: Devesh Walawalkar, Btech
Veermata Jijabai Technological Institute
Mumbai, Maharashtra INDIA
Corresponding Author Secondary
Information:
Corresponding Author's Institution: Veermata Jijabai Technological Institute
Corresponding Author's Secondary
Institution:
First Author: Devesh Walawalkar, Btech
First Author Secondary Information:
Order of Authors: Devesh Walawalkar, Btech
Order of Authors Secondary Information:
Funding Information:
Abstract: Early and correct diagnosis is a very important aspect of cancer treatment. Detection of
tumour in Computed Tomography scan is a tedious and tricky task which requires
expert knowledge and a lot of human working hours. As small human error is present
in any work he does, it is possible that a CT scan could be misdiagnosed causing the
patient to become terminal. This paper introduces a novel fully automated framework
which helps to detect and segment tumour, if present in a lung CT scan series. It also
provides useful analysis of the detected tumour such as its approximate volume, centre
location and more. The framework provides a single click solution which analyzes all
CT images of a single patient series in one go. It helps to reduce the work of manually
going through each CT slice and provides quicker and more accurate tumour
diagnosis. It makes use of customized image processing and image segmentation
methods, to detect and segment the prospective tumour region from the CT scan. It
then uses a trained ensemble classifier to correctly classify the segmented region as
being tumour or not. Tumour analysis further computed can then be used to determine
malignity of the tumour. With an accuracy of 98.14%, the implemented framework can
be used in various practical scenarios, capable of eliminating need of any expert
pathologist intervention.
Suggested Reviewers: Anthony Joseph Yezzi, B.Tech,Ph.D
Georgia Institute of Technology
anthony.yezzi@ece.gatech.edu
His research interests primarily lie in Image processing and computer vision,
particularly medical image analysis. His previous work includes medical imaging
related to MRI and CT scans. Hence I feel he is an ideal reviewer for my submitted
paper.
Powered by Editorial Manager® and ProduXion Manager® from Aries Systems Corporation
Journal of Machine Vision and Applications manuscript No.
(will be inserted by the editor)
A fully automated framework for lung tumour detection,
segmentation and analysis
Devesh Walawalkar
Received: date / Accepted: date
Abstract Early and correct diagnosis is a very impor-
tant aspect of cancer treatment. Detection of tumour
in Computed Tomography scan is a tedious and tricky
task which requires expert knowledge and a lot of hu-
man working hours. As small human error is present
in any work he does, it is possible that a CT scan
could be misdiagnosed causing the patient to become
terminal. This paper introduces a novel fully automated
framework which helps to detect and segment tumour,
if present in a lung CT scan series. It also provides use-
ful analysis of the detected tumour such as its approxi-
mate volume, centre location and more. The framework
provides a single click solution which analyzes all CT
images of a single patient series in one go. It helps to
reduce the work of manually going through each CT
slice and provides quicker and more accurate tumour
diagnosis. It makes use of customized image processing
and image segmentation methods, to detect and seg-
ment the prospective tumour region from the CT scan.
It then uses a trained ensemble classifier to correctly
classify the segmented region as being tumour or not.
Tumour analysis further computed can then be used to
determine malignity of the tumour. With an accuracy
of 98.14%, the implemented framework can be used in
various practical scenarios, capable of eliminating need
of any expert pathologist intervention.
Keywords Biomedical Image Analysis · Image
Segmentation · Computer Vision · Marker-controlled
Watershed Algorithm · Ensemble learning classification
Devesh Walawalkar
Bachelor of Technology in Electronics from V.J.T.I.,Mumbai
1005,11th floor,Hrishikesh Apts.,
Veer Savarkar Marg,Dadar(W),Mumbai,India-400028
Tel.: +919820143154
E-mail: devwalkar64@gmail.com
ORCID:0000-0001-9464-9027
1 Introduction
According to American cancer society, about 10-20%
of all cancer patients (approximately 1.7 millions) are
misdiagnosed every year and that at least 40,000 of
these patients die just because of it. Moreover, cancer is
the second largest cause of death in United states [20].
When a pathologist has on an average 80 Computed To-
mography (CT) slides per patient to be analyzed, there
is a likelihood of misdiagnosing one. Hence there is a
need to automate the tumour detection process and in
turn eliminate the human error that might creep in.
There is potential for lung cancer to be diagnosed at
an earlier stage through the use of screening with low-
dose computed tomography, which has been shown to
reduce lung cancer mortality by up to 20% [1,15]. Lung
CT scan diagnosis is however a highly specialized task
requiring expert knowledge. It is quite tricky and time
consuming task when done manually. In false negative
cases, the tumour manages to miss the human eye as
they can be quite similar in texture and size to other
particles present in the lungs.
Present work in this topic mainly focuses on qualitative
analysis of tumour discovered manually by a pathologist
in a single CT scan [2,11,18,27]. The literature lacks a
fully automatic framework which provides automatic
tumour detection of an entire CT scan series in a single
run, prioritizing framework’s practicability. Variance in
tumour’s size, its shape and location limits accuracy
of Image segmentation and other methods applied in
present work. What distinguishes the proposed segmen-
tation method from present work is the pre-processing
done on these images, which greatly helps to improve
its accuracy.
Further this paper proposes novel features extracted
from the segmented region, which help to distinguish
Manuscript latex file Click here to download Manuscript template.tex
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
2 Devesh Walawalkar
positive tumour cases from the negative ones. This pa-
per aims to provide a practical solution in order to re-
duce the time and effort required in manual detection
process, achieve high accuracy and to completely elim-
inate human error from the diagnosis.
2 Database Collection
The incorporated database consists of CT scan series of
61 patients, each with tumour present in one or more
slices. Each series consists of about 70 to 80 CT images.
The Database is obtained from the Cancer Imaging
Archive (TCIA) public access network [5]. This archive
contains collections of CT, PET, MRI scan series of can-
cer patients, of various body organs like lungs, brain,
bladder, prostate etc. This collection is freely available
to browse, download, and use for commercial, scientific
and educational purposes as outlined in ‘Creative Com-
mons Attribution 3.0 Unported License’. The database
specifically named ‘Lung CT diagnosis’ was created at
Moffitt Cancer Centre (Tampa, Florida) [8].
All images are diagnostic contrast enhanced CT scans.
The images were retrospectively acquired, to ensure suf-
ficient patient follow-up. Slice thickness is variable be-
tween 3 and 6 mm. All images were taken at diagno-
sis and prior to surgery. Patient data was anonymized
and de-identified prior to the analysis. The database
includes tumour detection results prepared by experi-
enced radiologists. It provides individual slice numbers
of each patient series, where tumour was detected. Part
of these results are used to train the classifier incorpo-
rated in the framework and remaining are used to test
its accuracy.
Fig. 1 Framework user interface for inputting data and dis-
playing result analysis
3 Computing Environment
The entire framework is built using MATLAB software
package [14]. The testing of this framework and process-
ing of stated results were done on a computing system
having an Intel i5 fifth generation processor without
any special purpose GPU. With this configuration, the
framework took on an average about 1-1.5 minutes to
analyze a single patient CT series containing about 70
to 80 slices.
4 CT Scan Image Processing
The database contains respective CT images in DICOM
(Digital Imaging and Communication in Medicine) for-
mat. They are converted into 8 bit gray scale images
(Step 1 of Figure 3). The upper and lower 20% of the
image is blackened out as it does not contain any use-
ful information or a tumour. It could however mislead
the implemented segmentation algorithm as certain iso-
lated elements are present in it. The image is then mod-
ified such that pixels having intensity values in range of
110-130, which is the observed tumour intensity range
are stepped up to about 210-230 intensity range and
all other intensity ranges are stepped down to 10-30 in-
tensity range. This helps to highlight the segmentation
target i.e. tumour region in image. It helps the detec-
tion algorithm to provide correct and fine segmentation
of tumour region.
In some lung CT slices, the central part (mainly tra-
chea) of lungs appears isolated from its neighbouring
lung wall. This introduces a possibility of the part be-
ing falsely segmented by the algorithm. A vertical strip
consisting of intensity values in range of 10-30, is su-
perimposed along central region of image. This helps
to join central trachea part with its neighbouring wall
and correspondingly prevents any false segmentation.
An example of resulting processed image can be seen
in Step 2 of Figure 3. Closing operation [10,23] is per-
formed to fill in any gaps present inside the tumour.
This is followed by Opening operation [10,23] so as to
separate any possible tumour clinging loosely to lung
wall. The corresponding effect on image can be seen in
Step 3 of Figure 3.
5 Tumour segmentation
For segmentation of tumour, marker-controlled water-
shed algorithm [4,9,16,19,21,24,26,28] is implemented.
This method is selected on basis of the fact that lung
CT scan consists of a continuous connected mass with
only some small parts of it other than the tumour, be-
ing separated from the main body. Hence Image seg-
mentation is an ideal method to detect these isolated
tumours. Its implementation in medical image analysis
has been limited due to possibility of over-segmentation
and being noise sensitive. Processing the image to give
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
A fully automated framework for lung tumour detection, segmentation and analysis 3
Fig. 2 Entire framework flowchart
it apriori knowledge of which region we want to seg-
ment, significantly improves the algorithm efficiency.
Since watershed algorithm performs segmentation by
drawing ridge lines around local minimums in image,
we modify the image such that tumour and non-tumour
regions become local minimums. It is done as follows:
Firstly, ‘Opening by reconstruction’ technique and sub-
sequently ‘Closing by reconstruction’ technique [7,22] is
applied with ‘disk’ structuring element of size 8. Here,
a ‘disk’ shape is used as a mask as it somewhat re-
lates to tumour morphology. The size 8 was fixed after
experimentation with different sizes. This helps to flat-
ten out the regional maximums present inside image
objects pertaining to tumour and non-tumour parts.
Further foreground object marker (i.e. Tumour region)
is found out by finding regional maximums in image.
This process provides both foreground and background
markers (i.e. non-tumour region) present in image (Step
4 of Figure 3). They are distinguished using adaptive
thresholding technique [13]. The image is modified such
that uniform regional minimums are present at the loca-
tion of computed foreground and background markers
in the image. This in effect makes the tumour bound-
ary, a continuous regional maximum. This helps the wa-
tershed algorithm to draw ridge lines along tumour’s
boundary and hence segment it. Regional maximums
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
4 Devesh Walawalkar
are computed using ‘imregionalmax’ function present
in Image Processing toolbox [14] available in MATLAB
software.
Further a mask is created exactly identical to the shape
created by the watershed ridge lines (Steps 5 and 7 of
Figure 3). The segmented region is then extracted from
the image with help of this mask, by keeping all the
pixels inside and on the mask same and blacking out
the pixels outside the mask (Step 8 of Figure 3). The
MATLAB code for this mask generation can be found
in Online Resource 1.
6 Positive classification of Tumour
Classification of segmented region is necessary as cer-
tain small non-tumour particles (tertiary bronchi and
bronchioles) that are part of the lungs, appear isolated
inside certain CT slices. There is a possibility that the
implemented segmentation method could falsely seg-
ment this as a tumour. To avoid this, novel features
are computed from the extracted region. These include
the size (number of non-zero pixels) of the region, aver-
age intensity of region and pixel distance of region from
a vertical line passing through centre of image. The first
feature is computed simply by counting the number of
non-zero intensity pixels present in segmented region.
Second feature is calculated by dividing sum of in-
tensities of all non-zero pixels in the region by the first
computed feature i.e. size. This feature represents the
distinguishing property of tumour regions having a spe-
cific range of average intensity values only (between
100-120 intensity value), compared to non-tumour ones.
Third feature is computed by first finding centre
pixel of segmented region. Centre pixel is found out by
finding horizontal (X1 and X2) and vertical (Y1 and
Y2) extremities of region and then using equation (1).
Centrepixel = [(X1 + X2)/2, (Y 1 + Y 2)/2] (1)
Third feature is computed as the horizontal distance
of this centre pixel from vertical centre line in image.
Third feature represents the distinguishing property of
tumour region being fairly away from centre of im-
age compared to non-tumour ones. These features com-
bined together accurately help to distinguish tumour
from other non-tumour lung particles. Code for com-
puting these three features is provided in Online Re-
source 2.
A Bagged decision trees classifier [3,6,12,17,25,29]
is used here to correctly classify the segmented region.
It is implemented with help of Classification learner
application available in MATLAB software. Classifier
Table 1 Performance comparison of various classifiers on
cross validation set
Percent Accuracy
Classifier for N-fold cross validation set
N=5 N=10 N=15 N=20
Linear 90.24 91.17 92.13 90.73
Support Vector Machine
Decision Tree 89.21 89.53 89.49 89.79
K-Nearest Neighbours 92.39 93.02 90.64 92.11
(K=10)
Bagged Decision Trees 97.48 98.34 98.72 98.26
Gaussian 87.28 88.62 89.12 88.19
Support Vector Machine
Weighted 90.17 90.82 91.13 92.62
K-Nearest Neighbours
is trained using 15 fold cross validation technique on
60% of patient database. For training purpose, features
were computed and fed to the classifier with label ‘1’
for those containing tumour and label ‘0’ for those not
containing one. Labeling was done in accordance to
individual patient results provided with the database.
Bagged Decision trees is chosen over other classifica-
tion methods owing to its higher classification accuracy
(Table 1 and 2), greater control over its model fitting
parameters and its ensemble learning feature of using
multiple decision tree models having different subsets of
the database. This helps the model to better generalize
and reduce error further, compared to a single classifier
model[6,29].
Table 2 Performance comparison of various classifiers on
test set
Classifier Percent accuracy
for test set
Linear Support Vector Machine 87.96
Decision Tree 90.60
K-Nearest Neighbours (K=10) 94.21
Bagged Decision Trees 98.14
Gaussian Support Vector Machine 89.12
Weighted K-Nearest Neighbours 92.44
A performance comparison of various classifiers for
both cross validation set and test set is presented in
table 1 and 2 respectively.
7 Tumour analysis
Analysis of positively classified tumour region is further
carried out . It includes the maximum cross sectional
area of the tumour and its corresponding slice number,
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
A fully automated framework for lung tumour detection, segmentation and analysis 5
Fig. 3 Step wise processing of CT image while running through framework
its centre location within the image, approximate vol-
ume of the entire tumour and percent confidence of it
being a tumour (Figure 1). This analysis is computed
by finding out the upper and lower bounds of tumour
along both X, Y dimensions of the image and then by
using equation (1). Corresponding code can be found
in Online Resource 2. Cross sectional area is calculated
by counting the number of pixels representing the ex-
tracted tumour. Approximate volume is computed by
summing the cross sectional area of all positively com-
puted tumours and multiplying it by slice thickness of
that respective patient series. Maximum cross sectional
area is found by comparing areas of all positively de-
tected slices and finding maximum among it.
Table 3 Test set confusion matrix
Total slices Predicted Predicted Total
in test set Positive Negative
N = 1720
Actual 72 1 73
Positive
Actual 31 1616 1647
Negative
Total 103 1617 1720
8 Classifier trained model and Flowchart
A sequential flowchart of the proposed framework can
be seen in Figure 2. Classifier model used during frame-
work testing is provided in Online Resource 3. Users
simply need to add the file into the working directory
of MATLAB. The model named ‘Classifier’ could then
be loaded into their workspace.
The trained model can be used for prediction using ‘pre-
dict’ function provided by Statistics and Machine learn-
ing toolbox [14] present in MATLAB software package.
‘ESM 3.mat’ present in the provided resource is the
classifier pre-trained model. Users can use this trained
model or train one on their own using appropriate data
of segmented region features.
9 Results
Implemented framework gives an accuracy of 98.14%
on test set, for both the true positive and true neg-
ative cases combined. Considering only true positive
cases, accuracy is 98.63% and for true negative cases,
accuracy is 97.02% (Refer Table 3). Framework accu-
racy over entire database is 98.40%. The feature value
ranges for a positive tumour region learnt by classifier
are as follows: 100-120 intensity/pixel for average inten-
sity, 1000-5000 pixels for tumour size and 80-110 pixel
distance between tumour centre and vertical central line
in image. The step wise processing of CT scan image
through the framework can be seen in Figure 3. Indi-
vidual patient results for entire database can be found
in Online Resource 4.
10 Discussion
As seen from results, an accuracy of 98.14% is good
enough to implement this system in practical scenar-
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
6 Devesh Walawalkar
ios. The results also provide evidence for the fact that
the three novel tumour features proposed in this paper,
together help to classify segmented tumour from seg-
mented lung particles with excellent accuracy. The an-
alyzing speed of the framework at about 1-1.5 minutes
per series is a considerable improvement over manual
speed of an individual pathologist. Computed tumour
analysis would help the operating pathologist to gain
in depth knowledge of the detected tumour.
Furthermore, framework stores the positively detected
tumour slices in a run time created folder having the
patient ID as its name. An example of images stored
in these result folder can be seen in Step 6 of Figure 3.
This provides for the pathologist to review and confirm
the tumours detected by framework. This serves to add
an extra layer of positive tumour confirmation.
Certain limitations of this proposed framework also ex-
ist. It includes the fact that only CT images can be
processed. There is currently no option for PET and
MRI scans. It has been customized to detect tumours
in lung region only. Further modifications need to be
done for detection in other body organs. The classifier
accuracy is somewhat limited due to limited training
dataset. A much larger training dataset having greater
variance would make the classifier even more accurate
and reliable. Although present in very small percent
of the cases, framework shows lesser accuracy in cases
where the tumour is clinging onto the lung wall.
Acknowledgements Special thanks to Moffitt Cancer Cen-
tre (Tampa Florida, US) along with Ms Olya Stringfield, PhD
from the ‘Department of Cancer Imaging and Metabolism’
for preparing this database and for providing an indexed ta-
ble containing positive tumour slice numbers of each patient
series. Also special acknowledgment to The Cancer imaging
archive (TCIA) for making this database available under pub-
lic license and for its efficient maintenance.
References
1. Aberle, DR., Adams, AM., et al., National Lung Screening
Trial Research Team.: Reduced lung-cancer mortality with
low-dose computed tomographic screening. N Engl J Med.
2011;365:395-409.
2. Balagurunathan, Y., Gu, Y., Wang, H., Kumar, V.,
Grove, O., Hawkins,Gillies, R. J. et all (2014).: Repro-
ducibility and Prognosis of Quantitative Features Extracted
from CT Images. Translational Oncology, 7(1), 72-87.
DOI:10.1593/tlo.13844
3. Banfield, R. E., Hall, L. O., Bowyer, K. W., Kegelmeyer,
W. P. (2007).: A comparison of decision tree ensemble cre-
ation techniques. IEEE transactions on pattern analysis and
machine intelligence, 29(1), 173-180.
4. Beucher, S., Bilodeau, M.: ”Road segmentation and obsta-
cle detection by a fast watershed transformation,” Intelli-
gent Vehicles ’94 Symposium, Proceedings of the, 1994, pp.
296-301. DOI: 10.1109/IVS.1994.639531
5. Clark, K., Vendt, B., Smith, K., Freymann, J., Kirby, J.,
Koppel, P., Moore, S., Phillips, S., Maffitt, D., Pringle, M.,
Tarbox, L., Prior, F.: The Cancer Imaging Archive (TCIA):
Maintaining and Operating a Public Information Repos-
itory, Journal of Digital Imaging, Volume 26, Number 6,
December, 2013, pp 1045-1057.
6. Dietterich, T. G. (2000).: An experimental comparison of
three methods for constructing ensembles of decision trees:
Bagging, boosting, and randomization. Machine learning,
40(2), 139-157.
7. Gonzalez R., Woods R., Eddins S.: Digital Image pro-
cessing using MATLAB, Prentice-Hall, Inc. Upper Saddle
River, NJ, USA 2003
8. Grove, O., Berglund, A. E., Schabath, M. B., Aerts, H.
J. W. L., Dekker, A., Wang, H., Gillies, R. J. (2015,
March 4).: Quantitative Computed Tomographic Descrip-
tors Associate Tumor Shape Complexity and Intratumor
Heterogeneity with Prognosis in Lung Adenocarcinoma. (A.
Muoz-Barrutia, Ed.)PLOS ONE. Public Library of Science
(PLoS). http://doi.org/10.1371/journal.pone.0118261
9. Hamarneh, G., Li, X. (2009).: Watershed segmentation us-
ing prior shape and appearance knowledge. Image and Vi-
sion Computing, 27(1), 59-68.
10. Heijmans, H. J. A. M., Ronse, C.: The Algebraic Part
of Mathematical Morphology, Part One: Erosions and Di-
lations,Computer Vision, Graphics, and Image Processing,
vol. 50, pp. 245295, 1990.
11. Lambin, P., Rios-Velazquez, E., Leijenaar, R., Carvalho,
S., Stiphout, R. G., Granton, P., Aerts, H. J. (2012).: Ra-
diomics: Extracting more information from medical images
using advanced feature analysis. European Journal of Can-
cer, 48(4), 441-446. DOI:10.1016/j.ejca.2011.11.036
12. Lavanya, D., Rani, K. U. (2012).: Ensemble decision tree
classifier for breast cancer data. International Journal of
Information Technology Convergence and Services, 2(1), 17.
13. Lie, W. N. (1995).: Automatic target segmentation by
locally adaptive image thresholding. IEEE Transactions on
Image Processing, 4(7), 1036-1041.
14. MATLAB, Statistics and Machine Learning Toolbox, Im-
age Processing Toolbox and Classification Learner Appli-
cation Release 2015a, The MathWorks Inc., Natick, Mas-
sachusetts, United States
15. Marcus, PM., Doria-Rose, VP., Gareen, IF. et al.: Did
death certificates and a death review process agree on lung
cancer cause of death in the National Lung Screening Trial?
Clin Trials. 2016;13:434-438.
16. Mon, K. L. (2014).: Automatic Image Segmentation Us-
ing Edge and Marker-Controlled Watershed Transforma-
tion. In International Conference on Advances in Engineer-
ing and Technology (ICAET) March (pp. 29-30).
17. Oza, N. C., Russell, S. (2000, July).: Online ensemble
learning. In Aaai/Iaai (p. 1109).
18. Ravanelli, M., Farina, D., Morassi, M., Roca, E., Caval-
leri, G., Tassi, G., Maroldi, R. (2013).: Texture analysis of
advanced non-small cell lung cancer (NSCLC) on contrast-
enhanced computed tomography: prediction of the response
to the first-line chemotherapy. European Radiology, 23(12),
3450-3455. DOI:10.1007/s00330-013-2965-0
19. Ray, S., Hagge, R., Gillen, M., Cerejo, M., Shakeri, S.,
Beckett, L., Badawi, R. D. (2008).: Comparison of twodi-
mensional and threedimensional iterative watershed seg-
mentation methods in hepatic tumor volumetrics. Medical
physics, 35(12), 5869-5881.
20. Siegel, R. L., Miller, K. D., Jemal, A. (2017), Cancer
statistics, 2017. CA: A Cancer Journal for Clinicians, 67:
730. DOI:10.3322/caac.21387
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
A fully automated framework for lung tumour detection, segmentation and analysis 7
21. Tan, Y., Schwartz, L. H., Zhao, B. (2013).: Segmentation
of lung lesions on CT scans using watershed, active con-
tours, and Markov random field. Medical physics, 40(4).
22. Vincent, L.: ”Morphological grayscale reconstruction:
definition, efficient algorithm and applications in im-
age analysis,” Proceedings 1992 IEEE Computer So-
ciety Conference on Computer Vision and Pattern
Recognition, Champaign, IL, 1992, pp. 633-635. DOI:
10.1109/CVPR.1992.223122
23. Vincent, L. (1994): Morphological Area Openings and
Closings for Grey-scale Images. In: O, YL., Toet, A., Foster,
D., Heijmans, H.J.A.M., Meer, P. (eds) Shape in Picture.
NATO ASI Series (Series F: Computer and Systems Sci-
ences), vol 126. Springer, Berlin, Heidelberg
24. Yan, J., Zhao, B., Wang, L., Zelenetz, A., Schwartz, L.
H. (2006).: Markercontrolled watershed for lymphoma seg-
mentation in sequential CT images. Medical physics, 33(7),
2452-2460.
25. Zhang, C., Ma, Y. (2012).: Ensemble machine learning
(Vol. 1). New York: Springer.
26. Zhang, Y., Qu, H., Wang, Y. (2006, November).: Adap-
tive image segmentation based on fast thresholding and
image merging. In Artificial Reality and Telexistence–
Workshops, 2006. ICAT’06. 16th International Conference
on (pp. 308-311). IEEE.
27. Zhao, B., James, L. P., Moskowitz, C. S., Guo, P.,
Ginsberg, M. S., Lefkowitz, R. A., Schwartz, L. H.
(2009).: Evaluating Variability in Tumor Measurements
from Same-day Repeat CT Scans of Patients with Non-
Small Cell Lung Cancer. Radiology, 252(1), 263-272.
DOI:10.1148/radiol.2522081593
28. Zhou, X., Hayashi, T., Hara, T., Fujita, H., Yokoyama,
R., Kiryu, T., Hoshi, H. (2006).: Automatic segmentation
and recognition of anatomical lung structures from high-
resolution chest CT images. Computerized Medical Imaging
and Graphics, 30(5), 299-313.
29. Zhou, Z. H., Tang, W. (2003).: Selective ensemble of deci-
sion trees. Rough sets, fuzzy sets, data mining, and granular
computing, 589-589.
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
Manuscript latex file
Click here to access/download
Supplementary Material
template.aux
Manuscript latex file
Click here to access/download
Supplementary Material
template.blg
Manuscript latex file
Click here to access/download
Supplementary Material
template.log
Manuscript latex file
Click here to access/download
Supplementary Material
template.lot
Manuscript latex file
Click here to access/download
Supplementary Material
template.out
Manuscript latex file
Click here to access/download
Supplementary Material
template.tcp
Manuscript latex file
Click here to access/download
Supplementary Material
template.toc
Manuscript latex file
Click here to access/download
Supplementary Material
template.tps
Material 1
Click here to access/download
Supplementary Material
ESM_1.pdf
Material 2
Click here to access/download
Supplementary Material
ESM_2.pdf
Material 3
Click here to access/download
Supplementary Material
ESM_3.mat
Material 4
Click here to access/download
Supplementary Material
ESM_4.xlsx